Soft Prompt
Definition
A Soft Prompt controls a model with trainable numeric vectors instead of exclusively with written instructions.
A Soft Prompt is a trainable sequence of continuous vectors that is prepended as additional context to a frozen language model.
Simply explained
The model receives learned numeric cues prepended while its other settings remain unchanged.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
A team trains a few prompt vectors per classification task without changing all model weights.
Limits and pitfalls
- Soft Prompts are not directly human-readable; transferability and behavior outside the training data must be tested.
Origin & history
Prompt Tuning was described in 2021 as a parameter-efficient method for large language models.
Use in mid-sized companies
- Possible example: A team trains a few prompt vectors per classification task without changing all model weights.
Frequently asked questions
What does Continuous Prompt (Soft Prompt) mean?
A Soft Prompt is a trainable sequence of continuous vectors that is prepended as additional context to a frozen language model. What should you watch out for? Soft Prompts are not directly human-readable; transferability and behavior outside the training data must be tested.
Sources
- The Power of Scale for Parameter-Efficient Prompt Tuning · Lester et al., The Power of Scale for Parameter-Efficient Prompt Tuning
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: